VC legend Bill Gurley says companies staying private longer (5-7 years → 10-15) breaks VC math. No one—GP, LP, founder—has incentive to mark assets down, creating systemic mispricing. He warns of 'zombie unicorns' (overvalued private firms) and AI's 'gavage tube' capital race (forced massive funding, boom or bust). Key holdings: Stripe (founder may never IPO, uses secondary sales for liquidity), Databricks (pulled from IPO by huge offers), SpaceX (near trillion-dollar valuation, extreme case). Gurley is cautious, sees LP liquidity crisis as a catalyst for change.
Bill Gurley, in Episode 427 of Invest Like the Best, provides a comprehensive analysis of structural changes in the venture capital market. The core argument is that the prolonged period companies remain private has fundamentally undermined the mathematics of VC returns, creating a system lacking pr
Bill Gurley, former partner at Benchmark Capital, joins Invest Like the Best for the sixth time in this issue, offering a comprehensive analysis of the structural changes taking place in the venture capital market. The core judgment is: the significantly extended time companies remain private has undermined the math behind VC returns, and the entire system lacks proper incentives—GPs, LPs, and founders alike have no motivation to mark assets accurately, creating a systemic coordination problem. Important conclusions include: as a platform-level shift, assessing AI's revenue quality and international competitive dynamics is critical; the current IPO and M&A market is stagnant, a large number of "Zombie Unicorns" carry inflated valuations, and LPs face liquidity challenges. Gurley points out that the rules of the game in capital markets are being reshaped, and founders must adapt to this new reality.
Bill Gurley argues that the venture capital industry has a fundamental incentive problem—from GPs and LPs to founders, no party has the incentive to accurately mark the true value of private assets.
Gurley points out that in private markets (PE and VC), asset pricing is self-reported by GPs to LPs. Although auditors are involved, different institutions give vastly different valuations for the same asset. More critically, fund managers on the LP side are often compensated based on paper marks rather than actual realized returns, so they too lack the incentive to push GPs to mark down valuations. For founders, Gurley analyzes that while accurate pricing is beneficial for healthy company building in the long run, two major obstacles exist in reality: first, founders tend to view the highest valuation multiplied by their ownership stake as personal wealth, making a 70% markdown extremely painful; second, the existence of liquidation preference—when a company's valuation falls below cumulative funding, liquidation preference can take most of the proceeds in an acquisition, causing founders to lose the willingness to adjust valuations.
Gurley emphasizes that this is not about anyone intentionally acting maliciously, but rather a system-level coordination problem. He cites systems thinking theory, noting that "the behavior of a system is different from the behavior of its components." Each participant acts within their own rationality, yet collectively they lead to information distortion. He predicts that if this incentive structure remains unchanged, it will continue to distort capital allocation efficiency.
Gurley argues that the extension of companies' private tenure from 5-7 years to 10-15 years has a far more severe mathematical impact on VC returns than is superficially perceived.
He illustrates with specific numbers: assume an investment is expected to return $100 in 10 years. If delayed to 15 years, at a 10% annual cost of capital alone, it would need to return $160 to achieve the same return. But considering VC's risk cost (15-20% annualized) and annual employee equity dilution of 3-6%, the actual required return for this 5-year delay rises to $250. Gurley points to the fatal issue revealed by NVCA data—over the past five years, the proportion of capital returned to LPs by VC funds within the 5-10 year window has plummeted from the historical average of 20-30% to 5-7%.
Gurley further warns that the combination of time cost and dilution effects means that many highly anticipated "unicorns" may never generate returns commensurate with their risk. He emphasizes that this is not a simple IRR vs. DPI debate, but rather that the mathematical model of the entire asset class is breaking down. He laments: "People love to talk about fund returns after removing the big winners, but no one asks—what if you keep the big winner but all other investments go to zero? And that is exactly the direction we may be heading."
Gurley clearly affirms AI as a platform shift, but offers a sober examination of the investment logic behind it, with particular focus on revenue quality, capital race patterns, and the international competitive landscape.
Gurley notes that the revenue quality of the current AI industry is questionable. Many AI startups' revenue is essentially resale of compute—a layer wrapped on top of foundation models and hosting services, and may be in a negative gross margin state. He reminds investors that such revenue is double- or triple-counted in the system, and that unit economics will eventually matter. He cites observations from the Benchmark team: the key inflection point is when a company shifts from "experiment/sandbox mode" to "optimization mode"—when capital is abundant, companies can sustain a cash-burning model for a long time, but once they pivot to optimization, growth rates and revenue quality will face scrutiny.
Gurley emphasizes that the capital race (Gavage Tube) is the core operating logic of the current AI world. He creatively uses the metaphor of a "gavage tube" to describe capital being forcibly injected into companies, forcing all participants into a "home run or die" dynamic. He states bluntly: "Traditional company building does not burn $100-150 million a year, but all the big AI companies are doing it." He warns that this is repeating the mistakes of the zero-interest-rate era of 2021, just now dressed in the garb of an AI technology platform. For founders, he suggests that when forced to accept massive funding, they should at least cash out for themselves to reduce personal risk.
On international competition, Gurley believes that the most noteworthy development in China's AI field is not DeepSeek itself, but the chain reaction it triggered—major players like Alibaba, Xiaomi, and Baidu have all open-sourced their models, forming a competitive landscape of "four well-funded entities, all open source." He judges that this open-source ecosystem of mutual training and iteration will bring China a flexibility in experimentation and optionality that the US market cannot match. He hints that this could become one of the biggest challenges US AI companies face in global competition.
Gurley argues that the most critical variable in current capital markets is whether the LP liquidity crisis can become a catalyst for change, reshaping the rules of the entire system.
He points out that multiple pressures are emerging at the LP level: Yale, as the originator of the "Yale Model," is selling $6 billion in private equity in the market—a signal of historic significance. Gurley analyzes that the essence of the Yale Model was to first allocate heavily to illiquid assets, but when everyone imitates that strategy, its unique advantage disappears. He cites Howard Marks: "You make big money when you are non-consensus and right. But if everyone copies David Swensen (the Yale Model creator) and allocates 50% to illiquid assets, will it still work?" He judges that this may be the beginning of the consequences of US university endowments systematically over-allocating to private equity.
Gurley further notes that the liquidity squeeze at LPs is creating new market dynamics: university endowments are forced to issue bonds (US universities issued $12 billion in debt in Q1 2025, the third highest on record), cut research budgets (NIH/NSF indirect cost reductions), and these pressures will push LPs to adjust their allocation strategies to VC. He specifically highlights that the stance of Middle Eastern sovereign wealth funds could be a key variable—the CIO of the Qatar Investment Authority has publicly stated "the clock is ticking for private equity." If this view becomes consensus, it will greatly impact the currently ultra-loose capital supply.
Gurley is pessimistic about the system's self-correcting mechanism: "All components are currently self-reinforcing... Unless something happens at the LP level, I don't see a correction mechanism." He predicts that if a systemic pricing reset occurs, the opportunists will leave Silicon Valley, and true company builders will usher in a more efficient and authentic operating environment.
| Target | Guest Attitude | Key Data |
|---|---|---|
| Stripe | Not stated (discussed as a typical case) | Founder said may never go public; Benchmark provides liquidity through frequent trading; valuation approximately 1/10 to 1/3 of trillion-level |
| Databricks | Not stated (discussed as a typical case) | Before its impending IPO, was retained by Thrive and other institutions with an offer "too good to refuse", obtaining a relatively large equity stake |
| SpaceX | Not stated (discussed as an extreme case) | Valuation has a "one-third to one-third" path close to one trillion dollars (approximately $333 billion) |
| Anthropic | Not stated (as an AI winner case) | Listed by Gurley as one of the market-recognized AI leaders |
| OpenAI | Not stated (as an AI case) | Burns about $7 billion per year |
| Anduril | Not stated (as a hard tech case) | Valuation approximately $30 billion; recognized by the U.S. Department of Defense and actively selling |
| Sierra | Not stated (as an AI application case) | AI company led by Brett Taylor; Gurley expressed confidence in the authenticity of its business |
| Character.AI | Early risk warning (as a consumer AI case) | Early feedback shows "memory" and "voice" defects leading to insufficient user stickiness |
| DeepSeek | Neutral (as an international competition trigger) | After open-sourcing, triggered a chain reaction among Chinese AI companies |
| Alibaba (Qwen) | Neutral (as a China AI development) | Open-sourced the Qwen model |
| Xiaomi | Neutral (as a China AI development) | Launched an open-source model |
| Baidu | Neutral (as a China AI development) | Announced that its model will be open-sourced |
| ServiceNow | Not stated (as a large company AI adaptation case) | Website "extremely high in AI content"; considered to have quickly adapted to the AI platform shift |
| Microsoft | Not stated (as a large company AI adaptation case) | AI mentioned 67 times in earnings call; CEO Satya Nadella gave a two-hour speech on AI issues |
| Apple | Not stated (as a large company AI adaptation case) | Considered likely to respond to competition by acquiring AI companies (e.g., Perplexity) |
| Not stated (as a large company AI adaptation case) | Accused that "almost no one uses Gemini"; CEO Sundar Pichai admitted he had not read "The Innovator's Dilemma" | |
| Meta | Not stated (as a large company AI adaptation case) | Zuckerberg believes being public makes the company "operate more sharply" |
| Uber | Not stated (as a historical case) | Once engaged in a capital race with Lyft, Gurley experienced it personally |
| Lyft | Not stated (as a historical case) | The capital race with Uber was a key case in Gurley's experience |
| FTX | Not stated (as a negative case) | No board member questioned the founder's behavior, leading to disastrous consequences |
| Tesla | Not stated (as a hard tech success case) | Viewed as one of Elon Musk's exceptional success cases |
1. “No one is motivated to accurately mark prices” (Bill Gurley) — From GPs and LPs to founders, every participant acts rationally in self-interest, but the systemic consequence is persistent distortion of asset prices, becoming a “systemic coordination problem.”
2. “Gavage Tube capital competition” (Bill Gurley) — Capital is forcibly injected into companies, forcing all participants to accept massive funding rounds or be destroyed by competitors. Gurley uses this metaphor to describe the current AI industry’s “home run or death” competitive logic.
3. “The mathematical curse of time delay” (Bill Gurley) — With a 15–20% risk cost plus 5% annual dilution, a VC return requirement that is delayed by 5 years rises from $100 to $250. NVCA data confirms that the proportion of capital returned to LPs within a 5–10 year window has plummeted from 20–30% to 5–7%.
4. “Yale selling PE is the ultimate signal of model failure” (Bill Gurley) — Yale University, as the originator of the “Yale Model,” is selling $6 billion in private equity in the market, suggesting that when “everyone is doing illiquid assets,” the unique advantage of this strategy has disappeared.
5. “AI revenue may just be reselling computing power, and at negative gross margins” (Bill Gurley) — Many AI startups’ revenue is essentially “a layer wrapped around base models + hosting services,” double-counted three or four times, and unit economics will ultimately expose its fragility.
6. “China’s open-source AI competitive landscape may be more dangerous than that of the US” (Bill Gurley) — DeepSeek has triggered giants like Alibaba, Xiaomi, and Baidu to open source, forming an ecosystem of “four well-funded, all open-source” players that train and iterate on each other, bringing experimental flexibility that the US cannot match.
7. “Unit economics will eventually matter” (Bill Gurley) — Even if users can sell a model from two generations ago to customers at 1/100 of the price, companies must plan a transition from “experimentation mode” to “optimization mode,” or they will fail to achieve sustainable operations.
8. “Founders must learn how to lead; it’s not innate” (Bill Gurley) — Gurley cites Ben Horowitz’s passage about “founders needing to learn how to lead,” emphasizing that founders must actively learn the skills to lead large-scale organizations, not rely solely on talent.